Config Raises $27M Seed Round to Become the “TSMC of Robot Data”

Building a robot is hard. Teaching it to move is harder. Large language models can draw on the vast troves of text already available across the internet, but robot AI has no such shortcut. Every training sample — a reach, a grasp, an egg cracked open — has to be physically recorded. That takes robots, facilities, and the people to run them. It’s costly, and it doesn’t scale easily.

Config was built to solve that problem. The Seoul- and San Jose-based startup produces and supplies high-quality training data for robotic foundation model (RFM) developers. Rather than building robots of its own, Config bets on the infrastructure layer — positioning itself as the “TSMC of robotics.” Just as TSMC manufactures chips for Apple, Nvidia, and AMD without competing with any of them, Config wants to supply the data that powers every robot AI developer’s models, without stepping into the ring itself.

The company has raised $27 million in an oversubscribed seed round led by Samsung Venture Investment, at a valuation north of $200 million. Strategic investors include ZER01NE, the venture arm of Hyundai Motor Group, LG Technology Ventures, and SKT America, the U.S. VC unit of SK Telecom. Financial backers include Mirae Asset Venture Investment, Korea Development Bank, GS Futures, Kakao Ventures, and ZVC. Pieter Abbeel, co-founder of Covariant AI and a professor at UC Berkeley, also joined as an angel. The raise brings Config’s total funding to $34 million.

Roots in Meta and Twelve Labs

Config was co-founded in January 2025 by CEO Minjoon Seo, a former AI researcher at Meta and chief scientist at video understanding startup Twelve Labs. His four co-founders bring experience from Waymo, Google, and Naver.

Speaking with TechCrunch, Seo framed the data problem in blunt terms. Training an LLM means feeding it text that already exists on the internet. Training a robot means physically going out and collecting every data point — which requires the robot itself, a facility to run it in, and people to operate it. The result is that robot AI costs far more to develop than software-only AI, and those costs compound quickly as companies aim for more capable machines.

100,000 hours of motion data, collected in Hanoi and Seoul

Config runs data production operations across Seoul and Hanoi, Vietnam, employing nearly 300 people. It has accumulated over 100,000 hours of human motion data to date — more than 30 times the roughly 3,000 hours in AgiBot World, currently the largest comparable open-source dataset.

Raw volume, though, isn’t Config’s main pitch. The company’s core technology is about transforming human motion data into a form that actually suits the way robots move and interact with objects. Seo draws an analogy to language translation: you can’t expect a model trained on English materials to perform well on Korean tasks without conversion. The same logic applies here. “The data must be converted, not the model,” he said. “This conversion technology is Config’s core technical differentiator.”

Already generating revenue, gunning for $10M ARR

Co-founder and COO Jack Bang confirmed that Config is already bringing in revenue, with customers spanning large manufacturers, system integrators, and companies in agriculture and defense.

The fresh capital will be deployed across three fronts: expanding data collection in Vietnam and Seoul toward one million hours; scaling the enterprise platform to $10 million in ARR by end of 2026; and launching a cloud-based Robot-as-a-Service (RaaS) product that allows customers to run Config’s foundation model without any onboard hardware.

Standing apart from the model builders

The robotic foundation model space has attracted serious capital. Physical Intelligence raised $600 million led by Alphabet Capital G at a $5.6 billion valuation. Skild AI pulled in $1.4 billion from SoftBank at a $14 billion valuation. Genesis AI, backed by Khosla Ventures and Eclipse Ventures, recently revealed a proprietary robot hand and data collection glove. All three are building the models themselves. Config’s bet is that whoever wins that race will still need a reliable, scalable source of training data — and that’s the role Config intends to own.